SIP-21 - Spores

By: Heather Miller, Martin Odersky, and Philipp Haller

Updated September 15th, 2013

 

Functional programming languages are regularly touted as an enabling force, as an increasing number of applications become concurrent and distributed. However, managing closures in a concurrent or distributed environment, or writing APIs to be used by clients in such an environment, remains considerably precarious– complicated environments can be captured by these closures, which regularly leads to a whole host of potential hazards across libraries/frameworks in Scala’s standard library and its ecosystem.

Potential hazards when using closures incorrectly:

  • Memory leaks
  • Race conditions, due to capturing mutable references
  • Runtime serialization errors, due to unintended capture of references

This SIP outlines an abstraction, called spores, which enables safer use of closures in concurrent and distributed environments. This is achieved by controlling the environment which a spore can capture. Using an assignment-on-capture semantics, certain concurrency bugs due to capturing mutable references can be avoided.

Motivating Examples

Futures and Akka Actors

In the following example, an Akka actor spawns a future to concurrently process incoming requests.

Example 1:

def receive = {
  case Request(data) =>
    future {
      val result = transform(data)
      sender ! Response(result)
    }
}

Capturing sender in the above example is problematic, since it does not return a stable value. It is possible that the future’s body is executed at a time when the actor has started processing the next Request message which could be originating from a different actor. As a result, the Response message of the future might be sent to the wrong receiver.

Serialization

The following example uses Java Serialization to serialize a closure. However, serialization fails with a NotSerializableException due to the unintended capture of a reference to an enclosing object.

Example 2:

case class Helper(name: String)

class Main {
  val helper = Helper("the helper")

  val fun: Int => Unit = (x: Int) => {
    val result = x + " " + helper.toString
    println("The result is: " + result)
  }
}

Given the above class definitions, serializing the fun member of an instance of Main throws a NotSerializableException. This is unexpected, since fun refers only to serializable objects: x (an Int) and helper (an instance of a case class).

Here is an explanation of why the serialization of fun fails: since helper is a field, it is not actually copied when it is captured by the closure. Instead, when accessing helper its getter is invoked. This can be made explicit by replacing helper.toString by the invocation of its getter, this.helper.toString. Consequently, the fun closure captures this, not just a copy of helper. However, this is a reference to class Main which is not serializable.

The above example is not the only possible situation in which a closure can capture a reference to this or to an enclosing object in an unintended way. Thus, runtime errors when serializing closures are common.

Basic Usage

Spores have a few modes of usage. The simplest form is:

val s = spore {
  val h = helper
  (x: Int) => {
    val result = x + " " + h.toString
    println("The result is: " + result)
  }
}

In this example, no transformation is actually performed. Instead, the compiler simply ensures that the spore is well-formed, i.e., anything that’s captured is explicitly listed as a value definition before the spore’s closure. This ensures that the enclosing this instance is not accidentally captured, in this example.

Spores can also be used in for-comprehensions:

for { i <- collection
      j <- doSomething(i)
} yield s"${capture(i)}: result: $j"

Here, the fact that a spore is created is implicit, that is, the spore marker is not used explicitly. Spores come into play because the underlying map method of the type of doSomething(i) takes a spore as a parameter. The capture(i) syntax is an alternative way of declaring captured variables, in particular for use in for-comprehensions.

Finally, a regular function literal can be used as a spore. That is, a method that expects a spore can be passed a function literal so long as the function literal is well-formed.

def sendOverWire(s: Spore[Int, Int]): Unit = ...
sendOverWire((x: Int) => x * x - 2)

Design

The main idea behind spores is to provide an alternative way to create closure-like objects, in a way where the environment is controlled.

A spore is created as follows.

Example 3:

val s = spore {
  val h = helper
  (x: Int) => {
    val result = x + " " + h.toString
    println("The result is: " + result)
  }
}

The body of a spore consists of two parts:

  1. the spore header: a sequence of local value (val) declarations only, and
  2. the closure.

In general, a spore { ... } expression has the following shape.

Note that the value declarations described in point 1 above can be implicit but not lazy.

Figure 1:

spore {
  val x_1: T_1 = init_1
  ...
  val x_n: T_n = init_n
  (p_1: S_1, ..., p_m: S_m) => {
    <body>
  }
}

The types T_1, ..., T_n can also be inferred.

The closure of a spore has to satisfy the following rule. All free variables of the closure body have to be either

  1. parameters of the closure, or
  2. declared in the preceding sequence of local value declarations, or
  3. marked using capture (see corresponding section below).

Example 4:

case class Person(name: String, age: Int)
val outer1 = 0
val outer2 = Person("Jim", 35)
val s = spore {
  val inner = outer2
  (x: Int) => {
    s"The result is: ${x + inner.age + outer1}"
  }
}

In the above example, the spore’s closure is invalid, and would be rejected during compilation. The reason is that the variable outer1 is neither a parameter of the closure nor one of the spore’s value declarations (the only value declaration is: val inner = outer2).

Evaluation Semantics

In order to make the runtime behavior of a spore as intuitive as possible, the design leaves the evaluation semantics unchanged compared to regular closures. Basically, leaving out the spore marker results in a closure with the same runtime behavior.

For example,

spore {
  val l = this.logger
  () => new LoggingActor(l)
}

and

{
  val l = this.logger
  () => new LoggingActor(l)
}

have the same behavior at runtime. The rationale for this design decision is that the runtime behavior of closure-heavy code can already be hard to reason about. It would become even more difficult if we would introduce additional rules for spores.

Spore Type

The type of the spore is determined by the type and arity of the closure. If the closure has type A => B, then the spore has type Spore[A, B]. For convenience we also define spore types for two or more parameters.

In example 3, the type of s is Spore[Int, Unit]. Implementation The spore construct is a macro which

  • performs the checking described above, and which
  • replaces the spore body so that it creates an instance of one of the Spore traits, according to the arity of the closure of the spore.

The Spore trait for spores of arity 1 is declared as follows:

trait Spore[-T, +R] extends Function1[T, R]

For each function arity there exists a corresponding Spore trait of the same arity (called Spore2, Spore3, etc.)

Implicit Conversion

Regular function literals can be implicitly converted to spores. This implicit conversion has two benefits:

  1. it enables the use of spores in for-comprehensions.
  2. it makes the spore syntax more lightweight, which is important in frameworks such as Spark where users often create many small function literals.

This conversion is defined as a member of the Spore companion object, so it’s always in the implicit scope when passing a function literal as a method argument when a Spore is expected. For example, one can do the following:

def sendOverWire(s: Spore[Int, Int]): Unit = ...
sendOverWire((x: Int) => x * x - 2)

This is arguably much lighter-weight than having to declare a spore before passing it to sendOverWire.

In general, the implicit conversion will be successful if and only if the function literal is well-formed according to the spore rules (defined above in the Design section). Note that only function literals can be converted to spores. This is due to the fact that the body of the function literal has to be checked by the spore macro to make sure that the conversion is safe. For named function values (i.e., not literals) on the other hand, it’s not guaranteed that the function value’s body is available for the spore macro to check.

Capture Syntax and For-Comprehensions

To enable the use of spores with for-comprehensions, a capture syntax has been introduced to assist in the spore checking.

To see why this is necessary, let’s start with an example. Suppose we have a type for distributed collections:

trait DCollection[A] {
  def map[B](sp: Spore[A, B]): DCollection[B]
  def flatMap[B](sp: Spore[A, DCollection[B]]): DCollection[B]
}

This type, DCollection, might be implemented in a way where the data is distributed across machines in a cluster. Thus, the functions passed to map, flatMap, etc. have to be serializable. A simple way to ensure this is to require these arguments to be spores. However, we also would like for-comprehensions like the following to work:

def lookup(i: Int): DCollection[Int] = ...
val indices: DCollection[Int] = ...

for { i <- indices
      j <- lookup(i)
} yield j + i

A problem here is that the desugaring done by the compiler for for-comprehensions doesn’t know anything about spores. This is what the compiler produces from the above expression:

indices.flatMap(i => lookup(i).map(j => j + i))

The problem is that (j => j + i) is not a spore. Furthermore, making it a spore is not straightforward, as we can’t change the way for-comprehensions are translated.

We can overcome this by using the implicit conversion introduced in the previous section to convert the function literal implicitly to a spore.

However, in continuing to look at this example, it’s evident that the lambda still has the wrong shape. The captured variable i is not declared in the spore header (the list of value definitions preceding the closure within the spore), like a spore demands.

We can overcome this using the capture syntax – an alternative way of capturing paths. That is, instead of having to write:

{
  val captured = i
  j => j + i
}

One can also write:

(j => j + capture(i))

Thus, the above for-comprehension can be rewritten using spores and capture as follows:

for { i <- indices
      j <- lookup(i)
} yield j + capture(i)

Here, i is “captured” as it occurs syntactically after the arrow of another generator (it occurs after j <- lookup(i), the second generator in the for-comprehension).

Note: anything that is “captured” using capture may only be a path.

A path (as defined by the Scala Language Specification, section 3.1) is:

  • The empty path ε (which cannot be written explicitly in user programs).
  • C.this, where C references a class.
  • p.x where p is a path and x is a stable member of p.
  • C.super.x or C.super[M].x where C references a class and x references a stable member of the super class or designated parent class M of C.

The reason why captured expressions are restricted to paths is that otherwise the two closures

(x => <expr1> + capture(<expr2>))

and

(x => <expr1> + <expr2>)

(where <expr1> and <expr2> are not just paths) would not have the same runtime behavior, because in the first case, the closure would have to be transformed in a way that would evaluate <expr2> “outside of the closure”. Not only would this complicate the reasoning about spore-based code (see the section Evaluation Semantics above), but it’s not clear what “outside of the closure” even means in a context such as for-comprehensions.

Macro Expansion

An invocation of the spore macro expands the spore’s body as follows. Given the general shape of a spore as shown above, the spore macro produces the following code:

  new <spore implementation class>[S_1, ..., S_m, R]({
    val x_1: T_1 = init_1
    ...
    val x_n: T_n = init_n
    (p_1: S_1, ..., p_m: S_m) => {
      <body>
    }
  })

Note that, after checking, the spore macro need not do any further transformation, since implementation details such as unneeded remaining outer references are removed by the new backend intended for inclusion in Scala 2.11. It’s also useful to note that in some cases these unwanted outer references are already removed by the existing backend.

The spore implementation classes follow a simple pattern. For example, for arity 1, the implementation class is declared as follows:

  class SporeImpl[-T, +R](f: T => R) extends Spore[T, R] {
    def apply(x: T): R = f(x)
  }

Type Inference

Similar to regular functions and closures, the type of a spore should be inferred. Inferring the type of a spore amounts to inferring the type arguments when instantiating a spore implementation class:

  new <spore implementation class>[S_1, ..., S_m, R]({
    // ...
  })

In the above expression, the type arguments S_1, ..., S_m, and R should be inferred from the expected type.

Our current proposal is to solve this type inference problem in the context of the integration of Java SAM closures into Scala. Given that it is planned to eventually support such closures, and to support type inference for these closures as well, we plan to piggyback on the work done on type inference for SAMs in general to achieve type inference for spores.

Motivating Examples, Revisited

We now revisit the motivating examples we described in the above section, this time in the context of spores.

Futures and Akka actors

The safety of futures can be improved by requiring the body of a new future to be a nullary spore (a spore with an empty parameter list).

Using spores, example 1 can be re-written as follows:

def receive = {
  case Request(data) =>
    future(spore {
      val from = sender
      val d = data
      () => {
        val result = transform(d)
        from ! Response(result)
      }
    })
}

In this case, the problematic capturing of this is avoided, since the result of this.sender is assigned to the spore’s local value from when the spore is created. The spore conformity checking ensures that within the spore’s closure, only from and d are used.

Serialization

Using spores, example 2 can be re-written as follows:

case class Helper(name: String)

class Main {
  val helper = Helper("the helper")

  val fun: Spore[Int, Unit] = spore {
    val h = helper
    (x: Int) => {
      val result = x + " " + h.toString
      println("The result is: " + result)
    }
  }
}

Similar to example 1, the problematic capturing of this is avoided, since helper has to be assigned to a local value (here, h) so that it can be used inside the spore’s closure. As a result, fun can now be serialized without runtime errors, since h refers to a serializable object (a case class instance).

blog comments powered by Disqus

Contents